To install latest developmental version of wiseR in R
To launch the app:
Networks are one of the most intuitive representations of complex data. However, most networks rely on pair-wise associations which limits their use in making decisions. Bayesian Networks (BNs) are a class of probabilistic graphical models which can provide quantitative insights from data. These can be used both for probabilistic reasoning and causal inference depending upon the study design.
A BN is a directed acyclic graph and provides a single joint-multivariate fit on the data with a list of conditional independencies defining the model structure. Unlike multiple pair-wise measures of association, fitting a model decreases the chance of false edges because the structure has to agree with global and local distributions. Importantly, unlike most other forms of Artificial Intelligence and Machine Learning, BNs are not a black-box model. These are transparent, interpretable and help the user in reasoning about the data generative process by looking at the motifs. This can be immensely useful in learning systems such as healthcare where a feedback between the clinicians and the learning system is paramount for adoption. The structure (independencies) of a BN can be learnt directly from data using machine learning or be specified by the user, thus allowing expert knowledge to be injected. After the dependence structure within the data is learnt (or specified) the parameters are learnt on the network using one of many possible approaches. wiseR provides most of the existing approaches such as constraint based algorithms and score-based algorithms for parametrization of the Bayesian Network. Recommendations as per the state-of-the-art in the literature are specified at each step of the BN learning process (i.e. the recommendatin of score based algorithms over constraint based algorithms for learning structure and Bayesian Information Criteria over Akaike Information Criteria for evaluating the fit). Parametrization of the network enables predictions and inferences. These inferences can be purely observational (“seeing” the state of a variable and its effect on neighbours) or causal (“doing” something to a variable, i.e. interventions and observing the effect on downstream nodes). wiseR provides scoring methods both for observational (e.g. Bayesian Information Criterion) and interventional (e.g. modified Bayesian Dirichlet Equivalent score) datasets.
The flexibility provided by BNs in probabilistic reasoning and causal inference has made these one of the most widely used artificial intelligence methods in Computer Science. However, the sophistication required to code all the features has limited their use by domain experts outside of computer science, e.g. clinicians. wiseR plugs this gap by creating a GUI for the domain experts and is an end-to-end solution for learning, decision making and deploying decision tools as dashboards, all from the wiseR platform.
Carefully learnt BNs are representations of the generative process that produced the data. These generative processes are captured in the the form of three basic motifs present in the network: chain, fork and collider junctions.
This motif is the simplest structure in a BN with a sequence of nodes pointing in the same direction. The intermediate nodes are called mediators and conditioning on any of the mediators makes the flanking nodes independent of each other. A mediator can be thought of as a mechanism, hence capturing the information about mechanism removes the need for capturing the triggering process, hence making the triggering process independent of the outcome.
The second type of structure observed in a BN is a common parent pointing towards two or more child nodes. The parent represents the common cause and not accounting for the parent is a common source of error (confounding) in many statistical models built upon real world data. A folk example of such an effect is Age as a confounder (common cause) of IQ and shoe-size in children. Failure to condition upon age will lead to a spurious correlation between the IQ and shoe-size and change our reasoning (hence predictions) about the system being modeled.
These are one of the most interesting motifs in an Bayesian Network and are indicated by two nodes pointing towards a single child node. Although the two parent nodes are not causally connected, conditioning on the state of the common child opens up a ghost-path for probabilistic inference flow (inter-causal reasoning) which can explain many intriguing paradoxes (e.g. the Monty Hall problem).
Calling the wiseR() function from the wiseR package launches the application in the default browser. The recommended browser for wiseR is Chrome. On launching the application for the first time, it conducts background checks, which takes 5-10 seconds depending upon the machine. This delay only happens on the first launch and all subsequent launches take less than a couple of seconds to initialize the user interface.
The toolbar on the left side of the page has icons for navigating to home, analysis engine, github development version and team information at any point of time.
Click “Start Analyzing” for navigating to the analysis engine with tabs specifying the functions. Hovering over most of the tabs brings up a tool-tip that briefly describes the function performed on that action.
“Start Analyzing” takes us to the main page of the wiseR engine. An intuitive left-to-right ordering of tabs guides the user into the analysis. This page has 5 tabs named ‘App Settings’,‘Data’,‘Association Network’,‘Bayesian Network’ and ‘Publish your dashboard’, each tab covering a specific functionality (Figure S3). Each of these is described next.
This tab (Figure S3) is used to set the parallel computing option (number of cores) for learning BN structure. Learning structure is known to be NP-hard problem and may be time consuming on large datasets. Setting the number of clusters allows each core to learn a structure when bootstrap learning is performed. Since learning one structure cannot be parallelized, this is an example of task parallelization.
This tab is used to load, pre-process and explore the dataset (Figure S4.)
The dataset panel contains the upload and preprocess menu while explore panel is used to visualize the distribution of the data.
Users can upload their own data in one of the ‘.CSV’,‘.RData’ & ‘.txt’ formats. For large data, saving the data as .RData is recommended for its compression of data, hence efficient storage. For ‘.txt’ files different options of file formatting are provided such as:
New users trying to learn about BNs or testing the functionality of the app can work with the pre-loaded datasets (Alarm, Hailfinder, Asia, Coronary) to replicate BN analysis commonly demonstrated with these datasets.
Users must note that the default setting of the app is to treat all variables with less than 53 levels as factor variables. Variables with more than 53 levels are treated as numeric variables and the user is prompted for discretization of such numeric variables to factors. While this is a convenient general option, it may be true that a numeric variable may have less than 53 levels in a particular dataset. In such cases, the user can un-check this option and specify the numeric variables explicitly to be converted to factors in the next step.
Current limit of file upload is 8000 mb, however it is recommended to use .RData for large datasets.
The preprocess menu is used to edit the data before using through the app. It is essential that data is discrete and complete for learning association network and bayesian network. Multiple options to pre-process the data have been provided.
Users can download the pre-processed dataset in current state as a .CSV file
Pre-processed data can be explored by looking at their frequency distributions visualized as barplots. This is an important step in the analysis as it allows the users to catch any anomalies that may bias the downstream analysis.
Before going to Bayesian Network analysis, it is instructive to visualize the association graph constructed upon the data. Learning association networks on the data before bayesian learning can give useful insights. Additionally the association network can serve as a starting structure for Bayesian Network structure learning. This can be achieved by using the edgelist downloaded from association network analysis in the structure learning part of the app. Various association scores are provided in the app as explained in the next section.
The build menu is used to build association networks. It consists of different metrics for association strength which are used to build association networks including,
Each method returns a value between 0 and 1. 0 being two variables are not associated and 1 being highly associated. Users can select a threshold value between 0 and 1 for eliminating edges below the threshold.
Most real world networks have communities. These may form due to preferential attachment to hub nodes, often leading to a scale free network structure. To leverage this phenomenon, the app has a module/community detection pipeline revealing sub-graphs in the original graph. These graphs can be individually explored and analysed from the drop-down menu.
There are different clustering algorithms available in the app for module/community detection which is powered by linkcomm package, like ward.D (Recommended),ward.D2,single,complete,average,mcquitty, median,centroid for user to choose from.
Visualization is often key in deriving useful inferences from a complex dataset and these functions are provided through the “Visual Settings” menu.
Color/Shape attributes: The user can either select variable names or provide a vector of variable indices that need to be grouped for node sape and color. These tools are useful in analyzing bipartite or multi-partite networks.
n-Nearest Neighbor Visualization:To ease the exploration of large graphs a threshold for visible neighbor chain is provided, such that clicking on a node only display its neighbors upto that degree. Alternatively, a user can see the list of n-th degree neighbors by clicking on a variable return. This returns a list of nth degree neighbors. n can be interactively set.
Graph layout: Further, the app allows various graph layout options, some of which are more suitable for particular kinds of data. Useful layouts include tree (for directed), Sugiyama (for directed), star and circle (both for undirected graphs).
Font Size: User can adjust the font size of the node labels to improve visibility
Download graph:User also has the option to download the network graph in html format from the current tab or using the save network option present in the bottom right corner of the graph.
Association networks are essentially a list of edges with corresponding weights. This data obtained from association networks analysis can be exported as an edgelist in tabular form and can also be downloaded it as a .CSV file to allow user to switch between different networks analysis tools that can utilize edgelists.
This is the core functionality of the wiseR app. Multiple features have been added to state-of-the-art practices including bootstrapping, ensemble averaging, error bars of approximate inference etc. These ensure robustness of reasoning and decisions derived from the BN.
The structure of a graphical models reveals the generative process of the data at hand. For this reason, BNs are interpretable and intuitive. Learning the structure from data is the most challenging yet insightful step in a BN analysis. However, if the structure is already known, or can be guessed, this can be pre-specified by an expert. In this way, BNs represent a perfect hybrid of expert and data-driven learning.
This section relies on expert/prior knowledge and enables the user to upload a network graph as a .CSV file to initialize the bayesian learning graph. The user can also add, remove or reverse the direction of any arc, once the graph has been uploaded. (Only used in score-based learning)
In the absence of expert knowledge, independencies can be directly learnt from the data itself. The computational learning of structure can be done using a number of bayesian learning algorithms such as
Options used in structure learning include:-
Important Notes
The user does not have to do the bootstrap learning every time if they wish to explore different thresholds for edge and direction strength. This can be achieved by using the parameter tuning option.
The final learnt structure can be saved as a as a bnlearn object in .RData format through the save option.
A user returning to analysis should not need to repeat the above steps. An already learnt and saved bayesian network structure can be uploaded as an .RData file.
Options to upload both bootstrapped and direct structure and adjusting their parameters are available in the app.
wiseR allows expert checks at every step and an expert can identify edges and directions that are not plausible in the real world (e.g. smoking cannot influence natural gender, but the reverse can happen). An expert can add/remove arcs, or reverse the direction of the existing ones on the basis of plausibility.
wiseR allows user to work with an external graph through the app. It is possible user may want to test out newer graph learning algorithms which are not available through the app. This ensures the user access to app features provided they have a edgelist of the external graph in “from” and “to” structure saved as a “.csv”
With the current advances in Graph based machine learning, wiser provides the standard GNN based structure learning method through the app. The user is asked to provide the Python and Cuda environment path with all requisit libraries such as:
The user can provide the no. of bootstraps for averaged structure learning. Please note, the user must ensure CUDA enabled system for functioning of GNN through the App.
This section enables the analyst to validate the learnt structure using 10-fold cross-validation or hold-out testing.
The log-likelihood loss and network score outputs produced can then be used to judge the usability of the learned structure.
After the parametrization of BN is complete, it can be queried to derive inferences and reason about the data. The inference learning menu enables the user to learn and explore conditional probability plots on the learned structure. The user can set an event and multiple evidences can be inserted as per the query. This crucial feature enables the user to explore probability plots on event nodes in the network conditionalized on chain of evidences, which is the essence of decision making through bayesian networks.
Inferences can be learnt through two methods using wiseR
Assuming that the User is working with large datasets, approximate inference is the default setting. This can be switched in the app. It is observed that the inference plots produced using approximate inference with error bars, is almost as accurate as exact inference, hence can enable robust reasoning and decisions on large datasets where exact inference is not possible.
Apart from structure and inference learning which are the workhorse of a BN analysis, the bayesian network tab allows options for visualization and exploration
Heads-up: This section relies upon JAGS installation on their system. JAGS can be downloaded from its official page.
As per our knowledge, there is no available open-source tool or a package in R that integrates structure learning with decision making. Influence diagrams have typically been defined by hand by experts. wiseR integrates these two capabilities to help the analyst arrive at the best (or next-best) policy for optimizing a payoff. For this purpose, the use can create a payoff node based off the variable that they wish to optimize (known as the utility node, e.g. Blood Pressure optimization in the Intensive Care Unit). This brings up the table to enter the utility of each state in the range of -1 to 1. For example, a user may assign a negative utility to an abnormally low or high blood pressure event (-1) and a positive (+1) utitilty to normal blood pressure. Alternatively, if the user does not care as much about the high blood pressure as they would care about low blood pressure (e.g. shock in the ICU), they can assign a zero utility to the high blood pressure event. After this the user sets the decision nodes that are direct parents of the node their utility node. If the policy from parents is expected to be obvious, they can go to grandparents or orders higher, however the effects may diminish. Finally the user runs the Monte-carlo simulation that displays the best policy along with the payoffs in a sorted table. This leap from making subjective inferences to policy advice makes wiseR a white-box appproach to actionable, transparent and interpretable machine learning.
Finally, the user can abstract the learning and decision making process into their own dashboard deployments. We consider this section to be a key highlight of wiseR. Publishing the analysis as an interactive dashboard helps reproducibility and effective communication. wiseR enables the user to produce a custom dashboard of their results as an R package.
The user has options to select the name and theme for the dashboard, and select the build option. wiseR builds the dashboard as a downloadable R/Shiny package which can be served as a web-application or launched as a local Shiny app. The custom dashboard is equipped with the most useful features of wiseR such as the interactive graph exploration and inference learning and visualization.
This feature save author hassle of creating a dashboard for their findings, and helps propagating research output to the community in an efficient manner, thus helping bridging the gap between the research and actual use.
This section compares the features of wiseR with other existing Bayesian Networks tools (Table) Table Key:
| Name | OS | TA | MLE | SL | BL | INT | CV | PL | IL | CD | IC | CI | EI | AIE | DN | GNNL | PP | Free |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| wiseR | Yes (R) | Data Scientists,End users | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Bayesian Networks | Yes (R) | Data Scientists, End users | Yes | Yes | No | No | No | Yes | No | No | No | No | No | No | No | No | No | Yes |
| BayesiaLab | No | End Users | No | Yes | No | No | No | Yes | No | No | No | Yes | Yes | No | No | No | No | No |
| Bayes Server | No | End Users | No | Yes | No | No | No | Yes | No | No | No | Yes | Yes | No | No | No | No | No |
| BNJ | Yes (Java) | End Users | No | No | No | No | No | Yes | No | No | No | No | Yes | No | No | No | No | Yes |
| CIspace | Yes (Java) | End Users | No | No | No | No | No | No | No | No | No | No | Yes | No | No | No | No | Yes |
| OpenMarkov | Yes(Java) | End Users | No | Yes | No | No | No | Yes | No | No | No | Yes | Yes | No | No | No | No | Yes |
| Sam Iam | No | End Users | No | Yes | No | No | No | No | No | No | No | No | Yes | No | No | No | No | Yes |
| UnBBayes | Yes (Java) | End Users | No | Yes | No | No | No | Yes | No | No | No | Yes | Yes | No | No | No | No | Yes |
This section provides a hands-on example to conduct a step by step analysis to replicate one of the landmark results obtained using data-driven Bayesian Network learning and Inference, the Sachs dataset on yeast pathway modeling presented in Sachs et al. (2005), which utilizes bayesian learning on complex interventional data.
We upload the sachs interventional data which is in a space separated text file into the app. Variables are read as factors as it is a numerically encoded discrete data.
The variable INT holds interventional information, adjustment is made for the same in the preprocess menu. This will now be used as a part of bayesian structure learning.
Next we learn a simple bayesian network on the data using Hill climbing algorithm and mbde network score. This graph will now be used to initialize structure learning using tabu algorithm. This mechanism of graph initialization in structure learning is especially useful in score-based learning like Tabu, which prevents it from getting stuck in local maxima.
We now upload the graph as initialization for structure learning.
Next, we set the appropriate parameters for structure learning:
This performed a direct structure learning without bootstraps using tabu, to replicate the results as it was published in the study. However, we recommend that the user should select the bootstrap parameters and do a bootstrap learning, to produce more robust structures.
The final learnt structure replicates the results of the original paper and has all the validated arcs presented therein. It also discovered additional arcs which were not present in the original paper and may be useful for future research and exploration.
We now confirm exact inference results with approximate inference with error bar to show that the unique approximate inference mechanism provided in the app produces as accurate results as exact inference without computational constraints in case of large structures. For this purpose we set the node ‘Akt’ as event conditional on ‘Erk = 1’. In the probability plots produced we can clearly see that approximate inference with error bars produces as accurate results as exact inference within 0.05 range of error, which is negligible
This concludes the walk-through example for replicating a BN analysis using wiseR
This section provides a hands-on example to learn decision networks and policy optimization through wiser. We use the Alarm dataset available in the pre-loaded datasets of the app.
We uplaod the already discretized and complete Alarm dataset from the app
We learn the Network strcuture using the following parameters:
As shown in the network the BP (blood pressure) has 2 causal parents the CO(cardiac output) and TPR (Total Peripheral Resistance)
The aim of our decision network now is to find optimal values of CO and TPR to get a NORMAL value of BP. Thus we set our BP as utility node, payoffs of 1 for NORMAL, -0.5 for HIGH and -1 for LOW BP.
Now we set CO and TPR as the decision nodes in the network. We aim to learn an optimal policy for both to maximize payoff on utility node.
Finally from policy learning we can see that NORMAL values for both CO and TPR ensure maximum payoff and thus maximum probability of NORMAL BP. It is also interesting to note that a HIGH TPR more strongly affects high probability of NORMAL BP as compared to CO. Thus verifying the knowledge already available in medical domain and validaing the utility of wiser as a policy learning platform.
This concludes the walk-through example for replicating a Policy learning framework using Decison Netwroks using wiseR